Logistic regression of soft labeled instances via the evidential EM algorithm

Abstract

We address the issue of fitting a logistic regression model to soft labeled data, when the soft labels take the form of plausibility degrees for the classes. We propose to use the E2M algorithm to take this partial information into account. The resulting procedure iterates two steps: first, expected class memberships are computed using the soft labels and the current parameter estimates; then, new parameter estimates are obtained using these expected memberships. Experimental results show the interest of our approach when the data labels are corrupted with noise.